Diversified Portfolios in Continuous Time

Tomas Bj ork
Bertil N
aslund
Department of Finance
Stockholm School of Economics
Box 6501
S-113 83 Stockholm SWEDEN

Published in
European Finance Review
Vol. 1, 361-387. (1998)

Abstract

We study a financial market containing an infinite number of assets,

where each asset price is driven by an idiosyncratic random source as well
as by a systematic noise term. Introducing asymptotic assets which cor-
respond to certain infinitely well diversified portfolios we study absence of
(asymptotic) arbitrage, and in this context we obtain continuous time ex-
tensions of atemporal APT results. We also study completeness and deriva-
tive pricing, showing that the possibility of forming infinitely well diversified
portfolios has the property of completing the market. It also turns out that
models where the all risk is of diffusion type are qualitatively quite different
from models where one risk is of diffusion type and the other is of Pois-
son type. We also present a simple martingale based theory for absence of
asymptotic arbitrage.

The authors would like to thank David Lando and two anonymous referees for a
number of very helpful comments and suggestions.

1
Key words: Large economies, diversifiable risk, APT, asymptotic arbitrage, com-
pleteness, martingales.
JEL Classification: G12, G13,

1 Introduction
The main object of study in this paper is a financial market containing a
large number of traded assets. Each asset price is assumed to be driven by
a systematic random source as well as by an idiosyncratic source of random-
ness, and we will study certain well diversified self-financing portfolios in
this market. Our model is thus a continuous time extension of the classical
arbitrage pricing theory (APT) models studied in Ross (1976), Hubermann
(1982) and others. For other continuous time APT models see Chamberlein
(1988), Reisman (1992) and Back (1988). More recent studies are those of
Kabanov and Kramkov (1994,1996). We will sometimes compare our find-
ings with the classical CAPM results. For the continuous time CAPM, see
Merton (1973).
A typical project, in the classical theory as well as in the continuous time
models, has been to study the existence (or non-existence) of the asymptotic
arbitrage possibilities which may arise because of the existence of diversifiable
risk. In the atemporal studies this project has been carried out using geomet-
rical Hilbert space arguments, whereas most of the continuous time authors
use the modern theory of martingale measures. The deepest results in this
respect are obtained by Kabanov and Kramkov (1994,1996), who consider a
very general sequence of increasing markets and show how absence of asymp-
totic arbitrage is related to contiguity (in the sense of le Cam) properties of
certain sequences of probability measures.
It is notable, however, that pricing of derivatives, as well as completeness
questions, has hardly been studied at all within the framework of large mar-
kets. One of our main goals is to study precisely these questions.
The structure and aims of the present paper are as follows.

Every asset price is driven by one idiosyncratic source of randomness

as well by one systematic random factor. To make the exposition con-

2
 ceptually as clear as possible we have chosen to work within a very
simple model, where every random factor is either a Wiener or a Pois-
son process.

We introduce explicitely certain asymptotic assets into the model.

These assets have a natural interpretation as infinitely diversified port-
folios, and they correspond in reality to large mutual funds. Most of
the paper is then devoted to a study of the asymptotic model, which
consists of the original model plus the asymptotic assets.

By applying standard (finite) no arbitrage techniques to the asymptotic

model we easily obtain APT results for the various markets prices of
risk, and in particular we show that if the market is free of arbitrage
then the market price of diversifiable risk tends to zero.

In the most central part of the paper we study (asymptotic) complete-

ness, and pricing of derivatives within the asymptotic framework above.
The main result is that the inclusion of the asymptotic assets has the
effect of completing the market.

It turns out that, as regards hedging and pricing of derivatives, there is

a fundamental difference between, on the one hand models where both
the idiosyncratic and the diversifiable risk are of the same type (i.e.
Poisson-Poisson or Wiener-Wiener), and on the other hand the cases
when the diversifiable risk component is of a different type from from
the market risk (i.e. Poisson-Wiener or Wiener-Poisson).

In particular we show how price and hedge contingent claims in a jump-

diffusion model. It is to be noted that here the the asymptotic asset
plays a fundamental role in the hedging portfolio.

We end by presenting a martingale based theory of large markets, and

we study our model in the light of this theory.

3
2 The basic model

2.1 The market

We take as a priori given, a financial market containing a denumerable set of
traded assets, and in the sequel this market will be studied on a finite fixed
time interval [0, T ]. The price of asset number i, with i = 1, 2, is denoted
by Si and we assume that the dynamics of Si are given as follows, under an
objective probability measure P .

dSi (t) = i Si (t)dt + i Si (t)dZi(t) + i Si (t)dW (t). (1)

Here i , i and i are assumed to be known deterministic constants (see also

the technical assumptions below). The process W , which is common to all
assets, represents the systematic risk in the market and W is assumed to be
a standard P -Wiener process. The process Zi , on the other hand, represents
the asset specific risk of asset No i. We will consider two cases:

The case when Zi is a P -Wiener process (the diffusion case).

The case when Zi is a Poisson process with intensity i under the
measure P (the jump case).

We need some technical assumptions.

Assumption 2.1 We assume that

The time span under consideration is the a priori fixed interval [0, T ].
The parameters i , i , i and i , i = 1, 2, are uniformly bounded
deterministic constants.
The processes W, Z1 , Z2 , are P -independent.

We also assume the existence of a risk free asset with price process B, i.e.

dB = rBdt, (2)

where r is the deterministic short rate of interest.

4
Remark 2.1 The model above is chosen in order to be as simple as possible,
while still including the main ideas and leading to nontrivial results. It is of
course quite possible, and also natural, to consider more complicated models.
One can e.g. construct a model where each asset is driven by a finite number
of diversifiable factors, which are common only to a finite number of assets,
as well as by a finite number ( 2) of systematic factors which are common
to all assets. For reasons given above, we do not present such a general case
here, but in Section 5 we briefly present an extension to the case where, in
addition to the systematic diffusion factor, we also have a systematic jump
factor.

Remark 2.2 In the sequel we will often, for the sake of readability, suppress
the time index t. In order to discriminate between t and t we therefore
introduce the following notational convention for any process Y :

Y (t) = Y (t) = Y (t). (3)

2.2 Well diversified portfolios

We assume as above that the dynamics of asset No i are given, under the
objective measure P , by equation (1). We are allowed to form finite portfolios
in the various assets, and we will now see what happens when we form certain
infinitely well diversified portfolios. We need a small technical Lemma, and
we recall that i is the P intensity of Zi in the jump case.

Lemma 2.1 Define i by

i = i in the diffusion case and i = i i in the
jump case. Take the sequences {} i=1 , {}
i=1 , {}
i=1 and {}
i=1 as given.
Assume furthermore that there exists some positive constant A such that

sup |i | + sup |
i | + sup |i | A.
i i i

Then there exists at least one infinite increasing subsequence I = {ij }

j=1 of
the integers, and real numbers I ,
I and I (depending on the choice of the
subsequence), such that

lim i = I , lim
i =
I , lim i = I .
iI iI iI

5
Proof. Choose a subsequence I0 such that limiI i = lim supi i . Then
choose a subsequence I1 I0 such that limiI1
i = lim supiI0
i , and at last
choose a subsequence I2 I1 such that limiI2 i = lim supiI1 i .

Definition 2.1 An index set (subsequence) I as in the Lemma above will

be called a converging index set. We introduce the natural equivalence
relation by writing I J if and only if I = J , I = J and I = J .
The family of (equivalence classes of ) converging index sets is denoted by C.

Observe that generically there will exist several converging index sets, and
that I is unique (i.e. C is a singleton) if and only if limi i , limi i and
limi i all exist, (where the limits are taken with respect to the set of natural
numbers).
We now turn to the construction of well diversified self financing portfolio
strategies. Let us therefore consider a fixed finite index set I = {i1 , ..., iK }
and recall that a self financed portfolio in the assets {Si ; i I} is specified
by

1. The initial capital V (0).

2. A sequence of predictable (and sufficiently integrable) stochastic processses
P
{ui; i I} such that iI ui = 1.

Remark 2.3 Unless otherwise specified, the filtration {Ft }t0 under consid-
eration will be defined by
Ft = {W (s), Zi (s); s t, i = 1, 2, } (4)
Thus the prefix predictable above, means (Ft )-predictable. In some cases
it will be of interest to consider portfolios which are adapted to a smaller
filtration than {Ft }t0 , and in such cases we will give separate remarks to
clarify the situation.

As usual ui(t) above is interpreted as the relative share of the total portfolio
value invested in asset No i at time t, and we also recall that the evolution
of the value process V is given by
X dSi
dV = V ui . (5)
iI Si

6
Definition 2.2 Consider a converging index set I = {ij }
j=1 , and define In
by In = {ij }j=1 . A diversifying strategy along I is defined a sequence
n

{un }
n=0 of self financing portfolios such that

1. For all i, n and t we have uni (t) 0.

2. The strategy un is based on the assets {Si ; i In }.
3. For all t have the following relation P -almost surely.
lim sup uni (t) = 0.
n iI ,t0
n

This definition formalizes the idea of a well diversified portfolio, in the sense
that each asset asymptotically contributes an infinitesimal part to the entire
portfolio, and the simplest concrete example is of course given by uni (t)
1/n. Let us now consider a fixed diversifying strategy along I, and let us
denote the value process corresponding to un by V n . We then have the
following portfolio dynamics.

X X X
dV n = V n uni i dt + Vn uni i dZi + V n uni i dW, (6)
iIn iIn iIn

and the immediate project is to study the behaviour of this equation as n

tends to infinity.
It turns out that the case of a diversifiable jump risk in some aspects is
qualitatively different from the case when the diversifiable risk is of diffusion
type. Therefore the two cases are treated separately.

3 The diffusion case

3.1 The asymptotic diffusion model

In the diffusion case the V -dynamics in (6) will be

X X X
dV = V
n n
uni i dt +V n
uni i dWi +V
n
uni i dW, (7)
iIn iIn iIn

7
where W, W1 , are independent P -Wiener processes. We now investigate
the asymptotic behaviour of the V n -sequence.
From the definition of a diversifying strategy it is easy to see that, as n tends
to infinity, we have the following results
X X
uni i I , uni i I . (8)
iIn iIn

This takes care of the first and third term in the right hand side of (7). For
the middle term we can look at the infinitesimal variance (recall that the
section is heuristic) to obtain

X X
V ar uni i dWi = (uni )2 i2 dt.
iIn iIn

Defining Bn by
Bn = sup |uni (t)|,
iIn ,t0

we have the trivial inequality

X X
(uni )2 i2 Bn uni i2 .
iIn iIn
P
As n tends to infinity iIn uni i2 I2 , whereas Bn 0. Thus we see that
in the limit the dynamical behaviour of the value process will be given by

dV = V I dt + V I dW. (9)

This equation gives us the asymptotic behaviour of an infinitely diversified

portfolio, and it is important to notice that the coefficients I and I depend
on the particular choice of converging index set I. We now introduce all
processes of the type (9) as formal asset prices in our model.

Definition 3.1 The asymptotic diffusion model consists of the following

price processes.

1. For each i the process Si with P -dynamics

dSi = i Si dt + i Si dWi + i Si dW, i = 1, 2, (10)

8
 2. For each converging index set I C, the process SI with P -dynamics
defined by
dSI = I SI dt + I SI dW, I C. (11)

3. The risk free asset price B with

dB = rBdt. (12)

The asymptotic model is thus an idelized picture of a financial market where

we are allowed to trade infinitely diversified portfolios. It is an idealization
in the same way that continuous trading and frictionless markets are
idealizations of real world phenomena. The real world counterparts to the
SI -processes are of course the mutual funds, which can contain hundreds of
different assets.

Remark 3.1 If we have only one systematic factor, and if this factor is
tradeable, then we get (10)-(11) without needing the asymptotic portfolio.
An example of a tradeable systematic factor is the price of oil, or rather the
percentage change in the price of oil as tested in Chen et al. (1986) with
mixed results.

3.2 No arbitrage and the market price of risk

We now go on to investigate the implications of no arbitrage in the asymptotic
diffusion model.

Assumption 3.1 We consider as given the asymptotic model (10)-(12).

This model is assumed to be free of arbitrage in the sense that no arbitrage
possibilities exist when we are allowed to trade in finite portfolios.

We have included the requirement of finite portfolios in order to be able to

use the classical theory of no arbitrage. Note however that since we now
are working with the asymptotic model, the assumption of no arbitrage is
really an assumption of no asymptotic arbitrage in terms of the original
model (1)-(2).

9
We can now apply the standard technique of constructing locally riskfree
portfolios in order to see what the implications are of the no arbitrage asump-
tion. We start by fixing two different converging index sets I, J and form a
portfolio based on SI and SJ . Then we choose portfolio weights in such a way
that the dW -term vanishes from the value process dynamics. Equating the
rate of return of this locally riskless portfolio with the short rate of interest
(the no arbitrage assumption) gives us the relation
I r J r
= . (13)
I J
Thus we infer the existence of a real number with the property that
I r
= , for all I C. (14)
I
We will refer to as the market price of systematic risk.
For each i = 1, 2, we then define the real number i , referred to as the
market price of type-i-risk, as the solution to the equation
i = r + i i + i . (15)
Summing up we have the following result.

Proposition 3.1 The market price of systematic risk, , and the market
price of risk of type i, i , are given by
I r
= , (16)
I
i r i I r
i = , (17)
i i I
where the equations above hold for all I C and all i = 1, 2, .

We note that and i above are (apart from a minus sign) precisely the
Girsanov kernels one would use naively in order to produce a martingale
measure Q with the property that the Q-dynamics of the asset prices would
have the form
i + i Si dW
dSi = rSi dt + i Si dW , (18)
dSI = rSI dt + I SI dW, (19)

10
where W and W i are independent Q-Wiener processes. The problem with
this approach is that if we thus make a Girsanov transformation for each
separate Wiener process, we have no guarantee that the measure Q that we
obtain is globally equivalent to the objective measure P . This problem will
be dealt with in the Section 6.

Remark 3.2 If we write (17) on the form

( )
1 i I
i = i r 2 (I r) , (20)
i I
we see that the market price for idiosyncratic risk can be written as the total
risk premium i r minus the risk premium associated with systematic risk,
 
dS dS
1 Cov Sii , SII
(I r) ,
I2 dt
 
dSi dSI
Cov ,S
1 Si
all divided by i . We also recognize the term I2
dt
I
as the traditional
beta of factor i.

Sometimes it is claimed that the market price of diversifiable risk has to

equal zero. Translated into our model a statement of this kind would thus
assert that i = 0 for all i , and we see at once from Proposition 3.1 that
this does not have to be the case. On the contrary, and quite in accordance
with atemporal APT, we see from (17) that the market price of an individual
diversifiable risk source can take any value. There is however an asymptotic
result, again extending a previous APT result, which supports the claim
made above by saying that, in the limit, the market price of diversifiable risk
has to approach zero.

Proposition 3.2

1. The relation
i = 0, (21)
holds for a particular i if and only if
i r I r
= , for all I C. (22)
i I

11
 2. We have the following limit result.

lim i = 0. (23)
i

Proof. The first part of the proposition follows immediately from (17). To
prove the second part we first note that, for any converging index set I, we
have
I r I I r
lim i = = 0.
iI I I I
Now take any infinite subsequence J of the natural numbers. It is easily seen
that J will contain some converging set I. Thus, using the result above, we
see that any subsequence of the full sequence {i }1 will contain a further
subsequence converging to zero. Hence the full sequence has to converge to
zero.

Remark 3.3 From Proposition 3.2 and (20) we also see that asymptotically
we are on the CAPM line
i I
i r = (I r)
I2

3.3 Hedging and completeness

In this section we will study how to price and how to hedge against a con-
tingent claim in the original model as well as in the asymptotic one. To
keep things simple we will assume that we have only one converging index
set I namely the full sequence I = {1, 2, }. This means that all coefficient
sequences converge, i.e. we have limi i = , limi i = and limi i = .
Furthermore we have only one asymptotic asset, which we will denote by S,
with price dynamics given by

dS = Sdt + SdW. (24)

We keep the assumption of only allowing our hedging portfolios containing

a finite number of assets (though for the asymptotic model we are again
allowed to trade in the asymptotic asset S).

12
We start by looking at the original model and the question is whether it is
complete or not. The answer to this question will of course depend on what
we allow as a contingent claim, and here there are several choices. We list
some of them.

Definition 3.2 Fix some point in time T . A (sufficiently integrable) sto-

chastic variable X is said to be a

1. contingent claim if

X {W (t), Wi (t); t T, i = 1, }

2. market observable contingent claim if

X {Si (t); t T, i = 1, }

3. finite contingent claim if there exist a number n such that

X {W (t), Wi (t); t T, i = 1, , n}

4. finite market observable contingent claim if there exists a number

n such that
X {Si (t); t T, i = 1, , n}

In this section we will only consider claims of finite type, and we have the
following results.

Proposition 3.3 The original model is incomplete with respect to the set of
finite claims. It is complete with respect to finite market observable contingent
claims.

Proof. Fix T and define X by X = W1 (T ). Then X is finite and it is easily

seen that X can not be replicated by a portfolio containing a finite number
of assets. Thus the original model is incomplete with respect to the set of
finite claims.

13
Consider on the other hand a market observable claim
X {Si (t); t T, i = 1, , n} .
n o
i ; i = 1, n by
Defining the Wiener processes W

i (t) = 1 [i Wi (t) + i W (t)] ,

W
di
with q
di = i2 + i2 ,
we can write the price dynamics for S1 , , Sn as
i,
dSi = i Si dt + di Si dW i = 1, , n.
i -processes to
Using e.g. Gram-Schmidt we can now orthogonalize the W
obtain
X
n
dSi = i Si dt + Si j,
Dij dW (25)
j=1
1, , W
where W n are independent Wiener processes and the diffusion matrix
D = [Dij ] is nonsingular. We are now back in a standard situation and
completeness follows.

Remark 3.4 In the previous proposition we note that, by the convention

of Remark 2.3, the portfolios are allowed to be adapted to the big filtra-
tion {W (s), Wi (s); s t, i = 1, , n}. For the completeness part of the
propsition, it is easily seen from the proof, that the replicating portfolio is in
fact adpted to the smaller filtration {Si (s); s t, i = 1, , n}.

Proposition 3.4 The asymptotic model is complete with respect to the set
of finite claims.

Proof. Suppose that X {W (t), Wi (t); t T, i = 1, , n}. We now

adjoin the asymptotic asset to the finite set of assets S1 , , Sn . Thus we
obtain the (n + 1)-dimensional asset price vector S, S1 , , Sn for which the
diffusion matrix clearly is nonsingular. Completeness now follows from stan-
dard results.

14
Remark 3.5 A priori, the replicating portfolio above is only Ft -adapted,
but a closer look reveals that it is in fact adapted to the filtration
{W (s), Wi (s); s t, i = 1, , n} or (equivalently) to the filtration
{S(s), Si (s); s t, i = 1, , n}.

The moral of the results above is that the possibility of using infinitely di-
versifiable portfolios has the effect of completing the market. For the present
case, where we have only Wiener processes as driving noise, this result should
however not be overemphazised and the reason is as follows.
The distinction between a (finite) general contingent claim and a market ob-
servable claim is indicated by the name: a market observable claim is actually
defined in terms of the various asset prices, which in our interpretation of the
model, can be observed on the market. A claim like Wi (T ) on the contrary,
is a claim that can never be paid out in a market where the available infor-
mation is generated by the prices only. The reason that it can not be paid
out is that, based upon price information only, it is impossible to determine
the value of X at time T .
Thus we may say that while the original model is incomplete (in a certain
sense), the incompleteness stems from the fact that we have overspecified
the number of driving Wiener processes. Except for certain pathological
claims, which will never occur in a concrete market situation, all (finite)
claims can in fact be replicated. We emphasize, however, that this phenom-
enon is highly dependent upon the fact that, in the present model, both the
asset specific risk and the systematic risk are of diffusion type. As we shall
see below the picture is radically changed when the asset specific risk is of
jump type.

3.4 Derivative pricing

Let us fix a finite market observable contingent claim X. From the preceeding
section we know that X can be replicated, so it has a unique arbitrage free
price process (t; X) given by (t; X) = V (t) where V is the value process
for the replicating portfolio. Using standard martingale arguments we may
of course also express the price as an expected value.

15
Proposition 3.5 For any square integrable finite market observable contin-
gent claim X {S1 , , Sn } the price process (t; X) is given by

(t; X) = er(T t) E Q [X| {S1 (s), , Sn (s); s t}] (26)

Here the Q-dynamics of (S1 , , Sn ) are given by

dSi = rSi dt + i Si dW , i = 1, , n,
i + i Si dW (27)

where W 1, , W
,W n are independent Q-Wiener processes.

As we already have noted, the price dynamics in (27) contains one redundant
Wiener process. For computational reasons it may therefore be advantageous
to reduce the model for (S1 , , Sn ) to a model containing exactly n Wiener
processes, as we did in (25). This will give us Q-dynamics of the form
X
n
dSi = rSi dt + Si Dij dWj? , i = 1, , n,
j=1

where Wn? , , Wn? are independent Wiener processes under Q. To give a

simple concrete example let X be a standard European call on S1 (T ), with
strike price K. Then we write can the Q-dynamics for S1 as
1,
dS1 = rS1 dt + di Si dW (28)
q
where d1 = 12 + 12 , and the value of the call is given by Black-Scholes
formula using the volatility d1 .
We thus see that for the pure diffusion case we have the following moral.
Moral: When it comes to pricing and hedging of derivatives defined in
terms of a finite number of underlying assets, the possibility of forming well
diversified portfolios plays no role whatsoever. In particular, the market
prices of diversifiable and asset specific risk do not have to be known in order
to price the derivative.
As we shall see below, the jump case presents a totally different picture.

16
4 The jump case

4.1 The jump model

In the jump case, the asset dynamics are given by

dSi (t) = i Si (t)dt + i Si (t)dNi (t) + i Si (t)dW (t). (29)

where N1 , N2 , are independent P -Poisson processes. We recall that the

constant P -intensity of Ni is denoted by i .
In this model, the Poisson process Ni represents the occurence of sudden
asset specific shocks. If i > 0 then the shocks will increase the value of
the stock, and this could be interpreted as a technological breakthrough or
suddenly improved market conditions. When i < 0, on the other hand,
the stock price will decrease at every jump time of the Poisson process,
representing negative shocks. The most striking example is perhaps the case
when i = 1, in which case the stock price will drop to zero at the fist jump
time of Ni . This has a very natural interpretation as a model of (complete)
default for asset No i, and in this context it would be natural to study various
credit risk derivatives. See Sections 4.4-4.5 below for a concrete worked out
example.
The first paper treating jump risk is the classic stock price model in Merton
(1976). There exists a large literature on jump risk in interest rate models,
where the focus often has been on credit risk. See Artzner and Delbaen
(1995), Duffie and Singleton (1995), Jarrow and Turnbull (1995), and Madan
and Unal (1997). For a fairly general theory of point process driven interest
rate models, see Bjork et al. (1997a,b).

4.2 The asymptotic jump model

Turning to well diversified portfolios, the V -dynamics in (6) will now have
the form

X X X
dV n = V n uni i dt + Vn uni i dNi + V n uni i dW, (30)
iIn iIn iIn

17
In order to facilitate the analysis it is convenient to write (30) on P -semimartingale
form as

X X
dV n = V n uni [i + i ] dt + Vn uni i [dNi i dt]
iIn iIn

X
+ Vn uni i dW, (31)
iIn

where as before
i is defined as

i = i i . (32)

We now let n tend to infinity, and as in Section 3.1 we have

X
i ] I +
uni [i + I , (33)
iIn
X
uni i I . (34)
iIn

For the middle term in (31) we see that the differential dNi i dt is a
martingale increment, and thus it has expected value zero. The infinitesimal
variance is given by
X X
V arP [ uni i [dNi i dt]] = (uni )2 i2 i dt.
iIn iIn
P
As in Section 3.1 it is easily seen that iIn (uni )2 i2 i 0 as n . Thus
the asymptotic behaviour of (31) is given by the equation

dV = (I +
I )V dt + V I dW, (35)

and we now define the asymptotic jump model.

Definition 4.1 The asymptotic jump model consists of the following

price processes.

1. For each i the process Si with P -dynamics

dSi = i Si dt + i Si dNi + i Si dW, i = 1, 2, (36)

18
 2. For each converging index set I C, the process SI with P -dynamics
defined by
dSI = (I + I )SI dt + I SI dW, I C. (37)

3. The risk free asset price B with

dB = rBdt. (38)

Remark 4.1 Note that in equation (36), the parameter i does not repre-
sent the systematic drift (under P ) of the Si process. More precisely: i is
the P -drift of Si between jumps. The overall drift of Si (under P ) is given by
i +
i , which can be seen from the P -semimartingale representation

dSi = Si (i + i i )dt + i Si [dNi i dt] + i Si dW, i = 1, 2, (39)

Since we will work under the martingale measure Q, as well as under P , we

will keep to the measure invariant formulation (36).

4.3 The market price of risk

As before we assume that the asymptotic model is free of arbitrage and we
now go on to find the market prices of risk for the diffusion part as well as for
the various jump parts. Again we perform the analysis the following naive
way:

Carry out a Girsanov transformation for the Wiener process using a

likelihood process of the form

dL0 = L0 ? dW,
L0 (0) = 1.

For each i we carry out a Girsanov transformation using a likelihood

process of the form

dLi = L i i [dNi i dt]

?

Li (0) = 1.

19
In this way we obtain a measure Q with the property that

Under Q the process W has the decomposition

,
dW = ? dt + dW
is Q-Wiener.
where W
The process Ni is a Q-Poisson process with Q-intensity Q
i given by

Q ?
i = (1 + i )i .

The asset price dynamics under Q are thus given by

i (1 + ?i )] Si dt+i Si [dNi (1 + ?i )i ]+i Si dW
dSi = [i + i ? + , (40)

dSI = {I +
I + I ? } SI dt + I SI dW
, I C. (41)
Now we want to choose the Girsanov kernels in such a way that all asset
prices have a local rate of return equal to the short rate of interest, i.e we
want to solve the following set of equations
i + i ? +
i (1 + ?i ) = r, (42)
I + I + I ? = r. (43)
This system is easily solved as
r I
I
? = , (44)
I
r i
i i
I
(r I
I )
?i = . (45)

i
At last we may define the market price of diffusion risk, , by
= ? (46)
and the market price of jump risk of type i, i , by
i = ?i . (47)

Again we see that absence of arbitrage generically does not imply that the
market price of diversifiable risk equals zero. Instead we have the following
results, the proof of which are identical to the proof for the Wiener case.

20
Proposition 4.1

1. The market price of jump risk of type i equals zero if and only if the
following relation hold for all I C.
i r
i + I r
I +
= . (48)
i I

2. We have the following asymptotic result

lim i = 0.
i

We thus see that the market price of jump risk of type i equals zero if and
only if the risk premium per unit of systematic volatility of asset i equals
the risk premium per unit of systematic volatility for any asymptotic asset.
Furthermore we see that if all assets are identical under P , i.e. the coefficients
for asset i does not depend on the index i, then in fact all market prices of
diversifiable risks equal zero.

Remark 4.2 We may rewrite (45) as

 
1
dS dS
1 Cov Sii , SII
i =
?
i r 2
i + I r)
(I +
i
I dt

and then argue as in Remark 3.2. Using Proposition 4.1 we see that we again
are asymptotically on the CAPM line
 
dS dS
1 Cov Sii , SII
i r = 2
i + I r)
(I +
I dt

As in the diffusion case we stress the fact that the arguments above are
partly heuristic, since there is no guarantee that the prospective martingale
measure Q obtained above is globally equivalent to the objective measure
P . This question will be handled in Section 6 below.

21
Remark 4.3 Contemplating the possible signs of the various market prices
of risk, we have from (46)-(47)
I r
I +
= ,
I
i r
i + i
I
(I I r)
+
i =
i
Assuming that i > 0 for all i, and thus implying I > 0, we then see that
the market price of diffusion risk, , is positive if and only if we have the
I r > 0, i.e. if and only if the asymptotic asset has a positive
relation I +
risk premium. This relation will of course hold in every risk averse market.
In the same way we see that the market price of jump risk of type i, i , is
positive if and only if we have the relation
i + i r I r
I +
>
i I
i.e. if and only if the risk premium per unit of systematic volatility is greater
for asset No i than for the asymptotic asset.

4.4 Completeness for the jump model

In order to study completeness we again have to specify the relevant classes
of contingent claims to be considered. To make things simple we assume that
the class of convergent index sets is a singleton, i.e. that all model coefficients
converge. Thus we only have one asymptotic asset, which will be denoted by
S, and its P -dynamics will be written as
dS = ( +
)Sdt + SdW.
We now come to the first major difference between the diffusion model and
the jump model. For the pure diffusion model the large sigma algebra
{W (t), Wi (t); t T, i = 1, , n} was strictly included in the market
observable sigma algebra {Si (t); t T, i = 1, , n}, which meant that
we had to introduce the concept of market observable claims. In the present
mixed case it is easy to see that the sigma algebras
{W (t), Ni (t); t T, i = 1, , n}

22
and
{Si (t); t T, i = 1, , n}
are equal. This motivates the following definitions.

Definition 4.2 Fix some point in time T . A (sufficiently integrable) sto-

chastic variable X is said to be a

1. contingent claim if X {W (t), Ni (t); t T, i = 1, }.

2. finite contingent claim if there exist a number n such that

X {W (t), Ni (t); t T, i = 1, , n}.

The main result concerning the original model is not surprising. Remember
that, both in the original, and in the asymptotic model, we are by assumption
only allowed to trade in finite portfolios.

Proposition 4.2 The original model is not even complete with repect to
finite (market observable) claims.

Proof. It is immediately clear that no claim claim of the form (N1 (T )),
where is any nonconstant function, can be replicated using a finite set of
assets.

Note that, in contrast with the pure diffusion case earlier, these claims are
in no way pathological. As an example consider the case when 1 = 1.
Then we have S1 (t) 0 for all t , where is the first jump time of N1 ,
and we can interpet as the time of default for asset 1. In this example it is
extremely natural to consider a claim of the form X = I {N1 (T ) = 0}, where
I is the indicator function, since such a claim acts as an insurance against
default. As we have seen there is no way of replicating such a claim in the
original model, but for the asymptotic model the situation is brighter.

Proposition 4.3 In the asymptotic model every (Q-square integrable) finite

claim can be replicated.

23
Proof. Suppose that the claim X is of the form
X {W (t), Ni (t); t T, i = 1, , n} .
Then we adjoin the asymptotic asset S to S1 , , Sn , and it is a stan-
dard exercise to see that X can be replicated using a portfolio based upon
S, S1 , , Sn .

Once again we thus see that the introduction of the asymptotic asset i.e.
the possibility of forming well diversified portfolios has the function of
completing the model (see the end of the next section for a simple concrete
example). We stress the fact that, in contrast to the pure diffusion case,
the completness of the asymptotic model really amounts to a substantial
improvement over the original model.

4.5 Pricing in the jump model

We now turn to the task of computing arbitrage free prices for (finite) deriv-
atives in the asymptotic jump model. The results are obvious given those of
the preceeding section.

Proposition 4.4 Assume that C is a singleton and consider any finite claim
X {W (t), Ni (t); t T, i = 1, , n} .
Then the arbitrage free price process, (t; X), of the claim is given by
(t; X) = er(T t) E Q [X| {S(u), Si (u); u t, i = 1, , n}] . (49)
Here the Q-dynamics are given by
,
dS = rSdt + SdW (50)
 
dSi = r iQ Si dt + i Si dNi + i Si dW
, i = 1, , n. (51)
where
iQ = i i (1 + ?i ), (52)
and where ?i is given by (45). Furthermore, the processes N1 , , Nn are
independent Q-Poisson processes, and the Q-intensity of Ni is given by
i = i (1 + i ) i = 1, n.
Q ?

24
For the case of a simple claim we may compute the price by solving a
difference-differential equation.

Proposition 4.5 Assume that the claim is of the form

X = (S1 (T ), , Sn (T )) , (53)

then the price process is given by (t; X) = F (t, S(t), S1 (t), , Sn (t)) where
F is the solution of the boundary value problem

F F Xn
F 1 2F 1Xn
2F
+ rs + (r iQ )si + 2 s2 2 + i2 s2i
t s 1 si 2 s 2 1 si 2
X
n
2F X
n
2F
+ i j si sj + i si s
i,j=1 si sj 1 si s
X n
+ Fi? rF = 0, (54)
1

with the boundary condition

F (T, s, s1 , , sn ) = (s1 , , sn ) . (55)

Here we have used the notation

Fi? (s, s1 , , sn ) = F (s, s1, , si (1 + i ), sn ) F (s, s1 , , si , sn ).

(56)

Proof. The Kolmogorov backward equation.

Example:
As an example to illustrate ideas, let us replicate the binary claim discussed
above. i.e. we consider a claim X of the form

X = I {N1 (T ) = 0} . (57)

25
We start by writing down the Q-dynamics of the discounted asset price
S(t) 1 (t)
processes Z(t) = B(t) and Z1 (t) = SB(t) . These are easily obtained from
(50)-(51) as

dZ = ZdW , (58)
dZ1 = 1 Z1 dN
1 + 1 Z1 dW
, (59)
1 is the Q-compensated N1 -process, defined as
where N
1 (t) = N1 (t) Q
N 1 t.

h  i
Next we define the Q-martingale M by M(t) = E Q erT I {N1 (T ) = 0} Ft ,
where Ft as usual is defined by (4). It is easily seen that we have
Q
M(t) = erT e1 (T t) I {N1 (t) = 0} ,

and that the stochastic differential of M is given by

dM = M(t)dN1 . (60)

It is well known, and easy to see, that if we can find Ft -predictable processes
h and h1 such that

h(t)dZ(t) + h1 (t)dZ1 (t) = dM(t), (61)

Then the claim can be replicated by a portfolio consisting, at each t, of h(t)

units of the asymptotic asset, h1 units of asset S1 and hB units of the risk
free asset, where hB is defined by

hB (t) = M(t) h(t)Z(t) h1 (t)Z1 (t) (62)

Plugging (58) and (59) into (61) and equating coefficients gives us

M(t)1
h(t) = (63)
1 Z(t)
M(t)
h1 (t) = , (64)
1 Z1 (t)

26
Thus the portfolio defined by (62), (63) and (64) will indeed replicate the
claim. Notice how the introduction of the well diversified portfolio S in this
case helps us to hedge a claim which, without the possibility of diversifying,
would have been unhedgeable. Also notice that the portfolio weights, which
a priori only are adapted to the filtration {Ft }t0 , actually are adapted to
the much smaller filtration generated by S1 and S.
The arbitrage free price process, (t; X) for the claim above is of course
given by the value process,V (t) of the replicating portfolio. It is again easily
seen that the value process is given by V (t) = ert M(t), so we have the pricing
formula Q
(t; X) = er(T t) e1 (T t) I {N1 (t) = 0} . (65)

5 A systematic jump component

In this section we will see how the framework above can be extended to
more general situations. Rather than doing this in complete generality, we
have chosen to study a simple extension of the jump model above, where
we, apart from the systematic diffusion factor, also include a systematic
jump component. The arguments are very similar to those of the previous
sections, so our discussion will be rather brief.

5.1 The model

The asset dynamics are given, under the objective measure P , by

dSi (t) = i Si (t)dt + i Si (t)dNi (t) + i Si (t)dW (t) + Si (t)i dN. (66)

where N, N1 , N2 , are independent P -Poisson processes. The sequence

{i }i is assumed to be uniformly bounded, and the concept of a converg-
ing index set is extended in the obvious way. The constant P -intensity of
N is denoted by . The novelty is the systematic component N, which rep-
resents market shocks, as opposed to Ni which represents an idiosyncratic
shock.

27
Arguing exactly as in Section 4.2, we obtain the following P -dynamics for a
well diversified portfolio along a converging index set.

I )V dt + V I dW + V I dN.
dV = (I + (67)

We can now present the corresponding asymptotic model.

Definition 5.1 The asymptotic model consists of the following price processes.

1. For each i the process Si with P -dynamics

dSi = i Si dt + i Si dNi + i Si dW + i Si dN, i = 1, 2, (68)

2. For each converging index set I C, the process SI with P -dynamics

defined by

I )SI dt + I SI dW + I SI dN,
dSI = (I + I C. (69)

5.2 The market price of risk

We follow the methodology and notation of Section 4.3, and change measure
from P to Q, while adding a Girsanov transformation of the form
dLN = L N [dN dt] ,
?

LN (0) = 1.

As before obtain a measure Q with the the properties of Section 4.3, and
with the added property that the process N is a Q-Poisson process with
Q-intensity Q given by
Q = (1 + ? ).

The connection between the Girsanov kernels ? ?i , ? , and the various

market prices of risk , i , is as before given by
= ? , = ? , i = ?i .

Arguing as in Section 4.3, and introducing the notation

i = i ,

28
we are led to the following set of equations for the determination of the
Girsanov kernels.

i (1 + ?i ) + i (1 + ? ) = r, i = 1, 2,
i + i ? + (70)
I + I ? + I (1 + ? ) = r, I C
I + (71)

We now come to the main difference between the present setup and the one
encountered earlier. In Section 4.3 the corresponding system of equations
(42)-(43), was shown to have a unique solution, but in our present setting
this is no longer necessarily the case. If e.g. the family C of converging
index sets, is a singleton, then there are infinitely many solutions to (70)-
(71). The economic reason is that in this case, the asymptotic model is
incomplete. In order to have completeness, we intuitively need one traded
asset for every source of randomness. We obviously have one asset for each
source of idiosyncratic randomness, and by adding the asymptotic asset we
also have one asset for one of the systematic jump factors. Since we now
have two systematic factors, we will however need two linearly independent
asymptotic assets in order to have completeness, and with a singleton C we
only have one. The formal expression of this intuitive reasoning is given as
follows.

Proposition 5.1 Assume that C contains two converging index sets I and
J with the property that the volatility matrix
!
I I
(72)
J J

is nonsingular. Then there is a unique solution to the system (70)-(71). In

other words, the market prices of risk are uniquely determined within the
asymptotic model.

Proof. Under the assumptions above, ? and ? are uniquely determined

by the system

I + I ? + I (1 + ? ) = r,
I +
J + J ? + J (1 + ? ) = r.
J +

29
The coefficient ?i is then determined by the equation

i + i ? +
i (1 + ?i ) + i (1 + ? ) = r.

Regardless of whether the market is complete or not, the market prices of

diversifiable risk will still asymptotically tend to zero.

Proposition 5.2 Assume that ?i , ? and ? is a solution of the system

(70)-(71). Then
lim ?i = 0.
i

Proof. Take any subsequence of the natural numbers. Then this sequence
will contain a convergent index set I. Letting i along I in equation
(70) shows that in fact ?i converges to a limit ?I , and we obtain

I + I ? +
I (1 + ?I ) + I (1 + ? ) = r.

From this equation, and from (71) (with the same I) we obtain ?I = 0.

5.3 Hedging and pricing

Given the arguments in the previous section, the completeness question is
already more or less resolved: In order to have a complete market we need
two linearly independent asymptotic assets.

Proposition 5.3 Assume that C contains two converging index sets I and
J with the property that the volatility matrix
!
I I
(73)
J J

is nonsingular. Then the asymptotic market is complete in the sense that

every (Q-square integrable) claim can be hedged.

30
Proof. Suppose that the claim X is of the form

X {W (t), N(t), Ni (t); t T, i = 1, , n} .

Then we adjoin the asymptotic assets SI and SJ to S1 , , Sn , and it is a

standard exercise to see that X can be replicated using a portfolio based
upon SI , SJ , S1 , , Sn .

Proposition 5.4 Consider any finite claim

X {W (t), Ni (t); t T, i = 1, , n} .

Then, under the assumptions of Proposition 5.3 the arbitrage free price process,
(t; X), of the claim is given by

(t; X) = er(T t) E Q [X| {SI (u), SJ (u) Si (u); u t, i = 1, , n}] .

(74)
Here the Q-dynamics are given by
 
dSI = + I S dN,
r IQ Sdt + I SdW (75)
 
dSJ = + J S dN,
r JQ Sdt + J SdW (76)
 
dSi = r iQ iQ Si dt + i Si dNi + i Si dW
+ i S dN. (77)

where

IQ = I (1 + ? ),
JQ = J (1 + ? ),
iQ = i i (1 + ?i ),

The processes N, N1 , , Nn are independent Q-Poisson processes, and the

Q-intensity of N and Ni is given by (1+? ) and Q ?
i = i (1+i ) respectively.

For the case of a simple claim of the form X = (S1 (T ), , Sn (T )) we

may as usual compute the arbitrage free price as the solution of a boundary
value problem. Since the arguments are exactly the same as in Section 4.5,
and since the expressions become quite messy, this is however left to the
interested reader.

31
6 The martingale approach
The strategy in the previous sections has roughly been the the following

We extended the original model to the asymptotic model by including

the set of well diversified portfolios as asymptotic assets.

We assumed that the asymptotic model was free of arbitrage in the

sense that no arbitrage opportunities existed for any finite submodel.

In this way we claimed that absence of (finite) arbitrage in the asymp-

totic model could be interpreted as absence of asymptotic arbitrage
in the original model.

A different approach would be to give a precise definition of asymptotic

arbitrage for the original model and then to investigate the implications of
absence of asymptotic arbitrage. A natural conjecture in this context is of
course that absence of asymptotic arbitrage is connected to the existence of
a martingale measure Q P , such that all discounted asset prices are
Q-martingales.
In this section we present a simple version of such a theory. We show that
the existence of a global martingale measure implies absence of arbitrage and
we also investigate the models above in the light of this theory.
A very deep study of asymptotic arbitrage has been carried out by Kabanov
and Kramkov (1994),(1996). Instead of studying the existence of globally
defined martingale measures, as we do below, they relate absence of asymp-
totic arbitrage to contiguity (in the sense of Le Cam) properties of certain
sequences of local martingale measures. Their theory is considerably more
general than the one presented here, but it also requires much harder math-
ematical tools. Our theory, which is a simple and straightforward general-
ization of the corresponding finite theory, is presented in the belief that it is
more easily understood by the general reader.
Let us denote the finite market consisting only of the assets S1 , , Sn by
Mn . We consider a fixed time horizon T . By hn we denote a self financing
trading strategy on Mn and the corresponding value process is denoted by

32
V n . There are several resonable definitions of asymptotic arbitrage, and the
simplest to use is perhaps the following.

Definition 6.1 An asymptotic arbitrage is a sequence of strategies {hn }

1
such that, for some real number c > 0, we have

1.
V n (t) c, t T, n

2.
V n (0) = 0, n.

3.
lim inf V n (T ) 0, P a.s.
n

4.  
n
P lim inf V (T ) > 0 > 0.
n

We use a straightforward definition of martingale measures.

Definition 6.2 A measure Q is called a martingale measure if the fol-

lowing conditions are satisfied.

1. P Q on {W (s), Ni (s); s T }

2. All discounted asset prices B 1 (t)Si (t) are Q-martingales.

Exactly as in the finite case we now have the following central (and extremely
easy) result.

Proposition 6.1 Assume that there exists a martingale measure Q. Then

there is no asymptotic aritrage.

33
Proof. Without loss of generality we may assume that the short rate
of interest equals zero. Suppose now that {hn } 1 actually realises an as-
n
ymptotic arbitrage. Then every V is a local martingale under Q, and
because of the requirement V n c we see that every V n is in fact a
Q-supermartingale. Thus we have 0 = V n (0) E Q [V n (T )], and Fatous
lemma gives us E Q [lim inf V n (T )] 0. However,the definition of an asymp-
totic arbitrage, plus the equivalence between P and Q implies the inequality
E P [lim inf V n (T )] > 0 which leads to a contradiction.

We will now study our earlier models in the light of Proposition 6.1, and since
the treatment of the diffusion model and the point process model are com-
pletely parallell we confine ourselves to a discussion of the diffusion model.
Viewing each Wn as the coordinate process on canonical space C[0, T ] we
N Q
may write = n=0 n , where n = C[0, T ], and P as P = n=0 Pn , where
Pn is Wiener measure.
Following Section3.2 we denote the (deterministic and constant) Girsanov
kernel for Pn by n for n > 0 and for n = 0. This will transform
Pn into the measure Qn On n , and the new measure Q on is defined by
Q
Q= 0 Qn . In order to make all discounted asset prices into Q-martingales,
we furthermore require that the kernels satisfy the relations

i = r + i i + i , i = 1, 2, (78)

To see if Q is a martingale measure according to the definition above it now

Q Q
only remains to check if n=0 Pn is equivalent to n=0 Qn . To this end we
use the Kakutani Dichotomy Theorem (see e.g. Durrett (1996), p.244) which
Q Q
says that n=0 Pn n=0 Qn if and only if the following condition holds:

Y q 
EP Ln > 0 (79)
n=0

where Ln = dQn /dPn on {Wn (s); s T }.

We immediately have
1 2
Ln = en Wn (T ) 2 n T ,

34
and the Kakutani condition (79) is easily seen to be equivalent to the condi-
tion X
2 + 2n < . (80)
1

Using (78), the condition (80) reads

X 1
2
+ 2
{n r n }2 < , (81)
n=0 n

which implies that

n r n (t) 0, (82)
and if we take the limit along any converging index set I we obtain
I r
(t) = , I C (83)
I

which, together with (78) gives us

n r n I r
n (t) = , I C. (84)
n i I

Thus we obtain exactly the same formulas as in Proposition 3.1. The dif-
ference between our different approaches is that the martingale approach is
more general than the asymptotic model appraoch. Concretely this is seen
by the fact that while we only obtain the result that n 0 in the analy-
P
sis using asymptotic assets, we have the stronger result 2
0 n < using
the martingale approach. The point of the asymptotic model approach is
of course that it gives an intuitive interpretation of the abstract martingale
results in terms of the well diversified portfolios

35
References
[1] Artzner, P. & Delbaen, F. (1995): Default risk insurance and incomplete
markets. Mathematical Finance 5, 187-195.

[2] Back, K. (1988): Factorization of returns in multiperiod securities mar-

kets. Working paper. School of Business, Indiana University.

[3] Bjork, T. & Kabanov, Y. & Runggaldier, W. (1997a): Bond market

structure in the presence of marked point processes. Mathematical Fi-
nance, 7, 211-239.

[4] Bjork, T. & Di Masi, G. & Kabanov, Y. & Runggaldier, W. (1997b):

Towards a general theory of bond markets.Finance and Stochastics, 1,
141-174.

[5] Bremaud, P. (1981) Point Processes and Queues: Martingale Dynamics.

Springer-Verlag, Berlin.

[6] Chamberlain, G.(1988) Asset pricing in multiperiod securities markets.

Econometrica. 56, 1283-1300.

[7] Chen, N. & Roll, R & Ross, S.A. (1986) Economic forces and the stock
market. journal of Business 59, 673-700.

[8] Duffie, D. & Singleton, K. (1995) Modeling of term structures of de-

faultable bonds. Working paper, graduate School of Business, Stanford
University.

[9] Durrett, R (1996) Probability: Theory and Examples (2:nd ed.).

Duxbury Press.

[10] Hubermann, G. (1982) A simple approach to Arbitrage Pricing Theory.

Journal of Economic Theory. 28, 183-191.

[11] Jacod, J. & Shiryaev A.N. (1987) Limit Theorems for Stochastic
Processes. Springer-Verlag, Berlin.

[12] Jarrow, R. & Turnbull. (1995) Pricing derivatives on financial securities

subject to credit risk. Journal of Finance 50, 53-86.

36
[13] Kabanov, Y.M. & Kramkov D.O. (1994) Large financial markets: as-
ymptotic arbitrage and contiguity. Probability Theory and its Applica-
tions. 39, 222-229.

[14] Kabanov, Y.M. & Kramkov D.O. (1996) Asymptotic arbitrage in large
financial markets. Preprint.

[15] Madan, D. & Unal, H. (1997)Pricing the risks of default. Working paper
College of Business and Managemant, University of Maryland.

[16] Merton, R. (1973) An intertemporal capital asset pricing model. Econo-

metrica. 41, 867-888.

[17] Merton, R. (1976) Option pricing when the underlying stock returns are
discontinuous. Journal of Financial Economics. 5, 125-144.

[18] Ross, S. (1976) The arbitrage theory of asset pricing. Journal of Eco-
nomic Theory. 13, 341-360.

[19] Reisman, H. (1992) Intertemporal Arbitrage Pricing Theory. The Review

of Financial Studies 5 No.1, 105-122.

37